Consumer-driven virtual power plants: A field experiment on the adoption and use of a prosocial technology

Page created by Cheryl Davis
 
CONTINUE READING
Consumer-driven virtual power plants: A field experiment on the adoption and use of a prosocial technology
Consumer-driven virtual power
plants: A field experiment on
the adoption and use of a
prosocial technology
Quentin Coutellier, Greer Gosnell, Zeynep Gürgüç,
Ralf Martin and Mirabelle Muûls

June 2021

Centre for Climate Change Economics
and Policy Working Paper No. 379
ISSN 2515-5709 (Online)

Grantham Research Institute on
Climate Change and the Environment
Working Paper No. 350
ISSN 2515-5717 (Online)
Consumer-driven virtual power plants: A field experiment on the adoption and use of a prosocial technology
The Centre for Climate Change Economics and Policy (CCCEP) was established by the University of Leeds and the London
School of Economics and Political Science in 2008 to advance public and private action on climate change through
innovative, rigorous research. The Centre is funded by the UK Economic and Social Research Council. Its third phase
started in October 2018 with seven projects:
         1. Low-carbon, climate-resilient cities
         2. Sustainable infrastructure finance
         3. Low-carbon industrial strategies in challenging contexts
         4. Integrating climate and development policies for ‘climate compatible development’
         5. Competitiveness in the low-carbon economy
         6. Incentives for behaviour change
         7. Climate information for adaptation

More information about CCCEP is available at www.cccep.ac.uk

The Grantham Research Institute on Climate Change and the Environment was established by the London School of
Economics and Political Science in 2008 to bring together international expertise on economics, finance, geography, the
environment, international development and political economy to create a world-leading centre for policy-relevant
research and training. The Institute is funded by the Grantham Foundation for the Protection of the Environment and a
number of other sources. It has 12 broad research areas:

         1.    Biodiversity
         2.    Climate change adaptation and resilience
         3.    Climate change governance, legislation and litigation
         4.    Environmental behaviour
         5.    Environmental economic theory
         6.    Environmental policy evaluation
         7.    International climate politics
         8.    Science and impacts of climate change
         9.    Sustainable finance
         10.   Sustainable natural resources
         11.   Transition to zero emissions growth
         12.   UK national and local climate policies

More information about the Grantham Research Institute is available at www.lse.ac.uk/GranthamInstitute

Suggested citation:
Coutellier Q, Gosnell G, Gürgüç Z, Martin R and Muûls M (2021) Consumer-driven virtual power plants: A field experiment
on the adoption and use of a prosocial technology. Centre for Climate Change Economics and Policy Working Paper
379/Grantham Research Institute on Climate Change and the Environment Working Paper 350. London: London School of
Economics and Political Science

This working paper is intended to stimulate discussion within the research community and among users of research, and its content may
have been submitted for publication in academic journals. It has been reviewed by at least one internal referee before publication. The
views expressed in this paper represent those of the authors and do not necessarily represent those of the host institutions or funders.
Consumer-Driven Virtual Power Plants: A Field
       Experiment on the Adoption and Use of a Prosocial
                          Technology

     Quentin Coutellier∗             Greer Gosnell†           Zeynep Gürgüç‡             Ralf Martin§
                                            Mirabelle Muûls¶

                                               June 2, 2021

                                                  Abstract

          Demand response strategies that rely on individual behavior change have consistently
      demonstrated that flexibility in energy demand exists, but that there are constraints to demand
      shifting such as limited attention or rebound effects. This paper analyzes residential energy
      demand flexibility through a series of field experiments on adoption and usage of WiFi-enabled
      smart plugs that largely overcome the need for home dwellers’ effort and attention altogether.
      With 144 active participants, we study the adoption and user interaction of devices for auto-
      mated dynamic energy demand-side management. Our design allows us to explore whether
      and how devices and incentive systems may be designed and deployed to improve load bal-
      ancing in the centralized energy grid. Such a system would reduce the need for inefficient and
      often carbon-intensive back-up power generators during periods of peak demand, and increase
      the possible share of energy supplied by intermittent renewable energy sources. Our results
      suggest that users are more inclined to participate if switch-off events are sufficiently long to
      provide a meaningful probability of winning lottery prizes, but that higher frequency of such
      events may reduce participation intensity. We also develop an algorithm to optimally schedule
      users’ power demand and switch-off events. Applying this to high resolution user and power
      market data suggests that both costs and implied CO2 impact could be reduced by nearly
      10% for the average user with more efficient scheduling of demand.

Keywords: Behavioral intervention, field experiment, household energy demand, demand side man-
agement, load balancing.

JEL codes: C93, 031, Q21, Q41, Q54, Q55

   ∗ Imperial College London; Department of Geography and the Environment, London School of Economics and

Political Science (LSE), London WC2A 2AE, UK
   † The Payne Institute for Public Policy, Colorado School of Mines, Golden, CO, 80401, USA; Grantham Research

Institute on Climate Change and the Environment, LSE. Corresponding author: ggosnell@mines.edu.
   ‡ Economics and Public Policy Department, Imperial College London, London SW7 2AZ, UK.
   § Economics and Public Policy Department, Imperial College London, London SW7 2AZ, UK; Centre for Eco-

nomic Performance, LSE and CEPR.
   ¶ Economics and Public Policy Department, Imperial College London; CEP, LSE. Imperial College Business

School, London SW7 2AZ, United Kingdom.

                                                      1
1     Introduction

Increasing attention to the threat posed by global climate change has underscored the need for
rapid development and deployment of innovative solutions to catalyze a sustainable low-carbon
energy transition. Policymakers’ and practitioners’ ability to reduce overall energy consumption
and manage energy market volatility is becoming increasingly important in this context. As de-
carbonization of electricity supply will rely on large-scale deployment of variable renewable energy
sources, smarter management of electricity consumption, i.e. demand-side management (DSM),
will prove essential to reach the global targets set out in the Paris Agreement (Grunewald and
Diakonova, 2018; Gielen et al., 2019).

Innovation and internet connectivity in electronic systems have opened the door for development of
advanced DSM solutions to accommodate fluctuations in energy supply and demand. Such devel-
opments expand energy suppliers’ and grid operators’ toolkits beyond supply-side load balancing
mechanisms that often themselves require continuous emissions to maintain ramping capabilities
of back-up power plants. Excessive energy demand can be shifted from peak to off-peak hours,
when electricity typically costs less, by switching off some appliances for limited periods of time.1
Thus, residential and commercial consumers have the ability—and, under a growing number of
dynamic retail energy plans, the incentive—to contribute to a more flexible energy grid (Carbon
Trust and Imperial College London, 2016).

Our research considers the extent to which energy consumers’ adoption of an inexpensive Internet
of Things (IoT)2 technology—specifically the smart plug—can allow for carbon and cost efficiency
improvements through increased flexibility in energy demand given an embedded stock of energy-
consuming goods and durables in the economy.3 We posit that scaled utilization of such technology
enables the formation of sizable and robust virtual power plants (VPPs). Frequently regarded as
important elements of a net-zero carbon future, VPPs aggregate a network of distributed (often
renewable) power sources, storage systems, and flexible energy consumers to optimize and dispatch
generation or consumption in a smart energy grid. Flexibility on the demand side is particularly
crucial during temporary periods of either high energy demand (i.e. peak load) or limited energy
supply from variable renewable sources.

On the supply side, technologies such as IoT applications and smart meters can reveal crucial infor-
mation about patterns and flexibility in households’ and businesses’ appliance-specific energy use.
Suppliers and grid operators can use this information—and, given access, use direct load control4 —
to minimize inefficiency and waste while balancing electricity demand with available supply, with
   1 Energy consumption from buildings comprises about a fifth of global energy consumption, with the proportion

expected to increase as residents of non-OECD countries adopt home electronics and appliances at increasing rates
(U.S. Energy Information Administration, 2019). In anticipation of growing energy demand from the residential
sector and increased electrification of cooking, heating, and cooling as well as automobiles (Ürge-Vorsatz et al.,
2015; Gielen et al., 2019), understanding and managing patterns of electricity use is becoming ever more crucial
to meeting global climate change mitigation objectives. Additionally, due to varying carbon intensity of the grid
throughout the day, the time of the day at which this growing electricity demand occurs will impact its carbon
footprint as well as grid reliability (Holland et al., 2016). Analysis of energy efficiency investments related to air
conditioning in California demonstrates that the value to society of energy efficiency upgrades increases when time
of use is taken into account (i.e. they are characterized by a “timing premium”), as peak savings occur during hours
of high value to electricity markets (Boomhower and Davis, 2020).
   2 According to the Oxford Dictionary, IoT refers to “the interconnection via the Internet of computing devices

embedded in everyday objects, enabling them to send and receive data.”
   3 Our research is pertinent to the literature that provides evidence on “load shifting” behavior (Borenstein et al.,

2002). Considering monetary incentives, recent experimental research supports the hypothesis that time-of-use
(TOU) electricity pricing, dynamic pricing, and price elasticity signals—that is, reminding consumers how much
electricity costs when it is being used—can strongly affect the timing of electricity consumption (Wolak, 2011; Jessoe
et al., 2014; Bradley et al., 2016; Harding and Sexton, 2017), though evidence across all demand response programs
is quite mixed (Parrish et al., 2019). Non-monetary incentives—which have been tested extensively in relation to
overall energy consumption (Abrahamse et al., 2005; Buckley, 2020)—have also been tested in the context of load
shifting (Alberts et al., 2016). Prest (2020) demonstrated that the effect of TOU electricity pricing, coupled with
price signals, leads to a 10% reduction in peak electricity usage, and information provision via in-home displays
augments these reductions to 15%. Typically, non-pecuniary messages significantly affect high users, while low users
are more responsive to financial incentives (List et al., 2017).
   4 Interventions using direct load control, e.g., through smart thermostats, have shown to reduce peak demand by

10% to over 80% of reference load (Ivanov et al., 2013; Parrish et al., 2019).

                                                          2
minimal intervention into consumers’ lifestyles. Furthermore, shaving peak load throughout the
year can have a significant impact on suppliers’ bottom lines, providing vast economic motivation
for utilities to provide consumers with incentives—to date, usually embedded in rate design—to
reduce consumption during these hours.5 Thus, smart devices simultaneously create opportunities
for cost savings and reduction in negative environmental externalities.

Widespread uptake of such technologies whose adoption creates benefits that extend beyond those
internalized by the adopter—what we call “prosocial technologies”—is crucial for clean energy
transitions. Prosocial technologies include, for example, residential solar photovoltaic panels or
energy-efficient appliances, and they are increasingly being developed to target a variety of social
aims, such as apps for tracing Covid-19 exposure, traffic warning devices, and vaccination trackers,
among others. This paper considers the adoption and usage of such a prosocial technology in the
residential energy domain.

When considering the potential of dynamic pricing in electricity, it has been recently assumed
that enabling automation is of crucial importance given the tendency of households to be inat-
tentive to electricity prices (Borenstein and Bushnell, 2019; Parrish et al., 2019), particularly over
time (Houde et al., 2013; Gilbert and Graff Zivin, 2014) and when payments lack salience (Sex-
ton, 2015).6 This paper contributes to the limited literature that tests this conjecture in the
field. IoT technologies—which create the potential to observe, disaggregate, and automate energy
consumption—are becoming increasingly pervasive across the economy, with billions of electron-
ics and appliances projected to be connected through IoT in the near future (Khajenasiri et al.,
2017). This rapid IoT deployment has brought to the fore a number of concerns regarding the
performance and reliability of the evolving smart energy system, as issues related to misuse, safety,
transparency, and data leakage shake consumers’ trust (Alaa et al., 2017; Nicholls et al., 2020;
Parrish et al., 2020). As a consequence, the automation that is potentially necessary for effective
implementation of demand response could be difficult to achieve.

We explore these issues by conducting a series of field experiments designed to explore the extent
to which user behavior might constrain what is technically feasible—for example, users may have
reservations when it comes to adopting “smart” or externally controlled devices—while also explor-
ing how users might be incentivized to overcome such reservations.7 Additionally, by observing
users’ tendencies to override our randomly timed smart plug switch-off events, we are able to gain
insight into true energy demand flexibility; that is, we can understand which times of day or days
of the week individuals may or may not be willing to turn off appliances.

We ran our experiments in university halls via an initiative called POWBAL (short for “power
balancing”)8 . To facilitate the program, we developed a web platform that collects data on how
much electricity is consumed through the plugs and allows us to remotely and briefly switch off
POWBAL-branded smart plugs, through which users could connect a time-flexible appliance or
electronic device. The program incentivized students to participate by offering them a free smart
plug and the possibility of earning monetary rewards commensurate with plug usage. We also
emphasized the prosocial element of participation by informing the user of the environmental
   5 As  detailed in Pratt and Erickson (2020), a utility’s demand during the annual peak hour determines the capacity
cost it will be required to pay in the following year to ensure the same amount is available to them during the next
annual peak hour (plus a reserve margin). This cost comprises a significant portion of a utility’s capacity costs,
which account for 25% of utilities’ total wholesale market expense, a percentage that is trending upward. Hence,
the ability to immediately shave peak load during particular hours of the year that are often difficult to forecast
can lead to dramatic cost savings that, if passed to the customer, could provide additional incentive for consumers
to accept direct load control mechanisms.
   6 Inattention severely impedes the benefits of dynamic pricing since active response to price changes is cognitively

costly. Automation paired with dynamic pricing aims to enable lower-cost responses to price changes that can
achieve at least three times larger reductions for households (Gillan and Gillan, 2017).
   7 In the context of climate change, any policy proposal to reduce emissions and energy demand will largely rely

on an appropriate combination of behavioral, technological, and institutional capabilities. While technologies and
institutions are often in the spotlight, influencing behavior is also crucial in tackling climate change (Stern, 2008).
For instance, understanding the adoption behavior surrounding relevant technologies—such as home insulation or
smart meters—is arguably as fundamental as developing the technology itself (Allcott and Mullainathan, 2010; Toft
et al., 2014; Bugden and Stedman, 2019; Hmielowski et al., 2019).
   8 See Appendix A Figure 12 for a picture of the POWBAL plug.

                                                          3
benefits of energy demand flexibility.9 Consenting participants each received a smart plug that
allowed our research team (as well as the participants themselves) to control the power to the
connected appliance remotely throughout the study period. Participants then accrued points in
proportion to the energy use we displaced, which in turn determined their likelihood of winning
monetary prizes in our fortnightly lotteries.

To gain insight into the role of energy service interruption in determining plug usage, participants
were randomly assigned to treatments characterized by differing duration and frequency of our
randomly-timed “switch-off events”: (i) up to 15 or 30 minutes at a time, and (ii) either once or
up to eight times a day. The higher risk of interrupted energy services in the ‘higher intensity’
treatments is compensated through increased probability of winning versatile digital gift cards in
the fortnightly lotteries. We assess this trade-off through differences in plug adoption and usage
conditional on treatment assignment.

From the perspective of internal validity, the chosen context (i.e. university student accom-
modation) allows us to largely disregard complex irrelevant factors—in particular energy cost
considerations—and focus on the question at hand. From an external validity perspective, the
university context may be limiting. On the other hand, we run the experiment at three different
campuses comprising diverse student populations, allowing for assessment of generalizability to
at least the UK student population. We centered our attention on a developed country context
due to (i) these countries’ extensive established centralized energy grid systems that are largely
sourced with emissions-intensive power generators (though may accommodate more renewable en-
ergy sources, provided flexibility), and (ii) the prevalence of energy-intensive home appliances and
electronics amenable to smart plug usage.10 Nevertheless, these features suggest immense po-
tential of such programs in the developing world—perhaps particularly in rapidly industrializing
economies—going forward, as per capita energy use seeks parity with the developed world and the
need for rapid energy transitions becomes increasingly urgent, in particular due to air quality and
climate concerns.

Our results suggest that users are more inclined to participate if switch-off events are sufficiently
long to provide a meaningful probability of winning lottery prizes, but that higher frequency of
such events may reduce participation intensity. However, individuals with relatively long but less
frequent switch-off events are also more inclined to override. We provide evidence that individuals
are motivated to consume more energy via the smart plug—contributing more to load balancing—
just after they have been randomly selected to receive a lottery reward.

We also develop a framework to optimally schedule users’ switch-off events and hence their power
demand. The algorithm takes into account high frequency power consumption data, users’ re-
sponses to switch-offs, and wholesale electricity market data on power prices and carbon content.
The framework suggests that for the average user, both implied carbon consumption and (whole-
sale) power costs could be reduced by 8%, and can be as high as 26% for some users.

The remainder of the paper is structured as follows. The next section describes the experimental
setup and data. The third section outlines our stylized model and presents our results. Section 4
concludes.
   9 Several studies on energy flexibility and curtailment have demonstrated the importance of prosocial motivations
(e.g., Asensio and Delmas, 2015; Pratt and Erickson, 2020).
  10 The energy demand displacement potential for large-scale systems of smart plugs in the residential sector

is economically significant in developed countries. For example, should plugs be deployed on 30% of domestic
refrigerators in the United Kingdom—each of which consumes between 150W to 400W on average—then a large
coal-fired power station (2.4GW) could be entirely displaced during peak hours.

                                                         4
2     Research Design

2.1     Sample Recruitment

We conducted a series of field experiments focused on assessing individuals’ willingness to adopt
POWBAL smart plugs and their subsequent energy consumption via these plugs under three
randomly assigned treatments. We recruited undergraduate and graduate students from five LSE
student halls, and from one student hall each at Imperial College London and University of Reading.
Eligibility was restricted to individuals living in single-occupancy rooms to ensure that usage of the
plugs derived from the sole individual exposed to treatment, and to maintain comparability in the
characteristics of the residences and residents within a given student hall. The invitation offered
participation in a research study aimed at designing systems to manage peak demand and smooth
household energy consumption, where the information provided to prospective participants varied
only in the frequency and duration with which they would experience switch-off events, as in Table
1 (see Appendix A for recruitment materials).

Our sample of prospective participants comprised 1798 residents, where 262, 648, and 888 occupants
from Imperial, LSE, and Reading (respectively) received an invitation to participate. From this
sample, 231 experimental subjects (30, 93, and 108, respectively) consented to participate in the
study and received plugs, of which 144 (21, 39, and 84, respectively) actively used the plugs.
We recruited participants via several channels, including email, phone app, flyers, in-person, or
welcome packs in their rooms upon move-in. We recruited a majority of our participants in person
inside the residence halls, using a predetermined script to describe the study and solicit consent.11

Students registered to participate in the study by completing a brief consent survey in the Qualtrics
survey software. The survey first asked individuals to read a one-page overview of the study’s
purpose and implications of participation, then check four boxes confirming that they had read the
information sheet, understood their participation rights, and were aware that their plugs could be
switched off for periods of time corresponding to their randomly assigned frequency and duration,
which were stated explicitly and saliently in both the information sheet and the consent language
(see Appendix A).

Using student halls as the context of study comes with advantages and drawbacks. Students are not
typical households and use much of their household electricity outside of their rooms, for example in
laundry or common dining rooms and kitchens. They therefore have a smaller range of appliances to
which they can choose to connect their plugs, with laptop and phone chargers comprising the bulk
of available electronics, followed by lamps and hair accessories, and only a handful of individuals
reporting larger energy consumers such as refrigerators, TVs, or electric heaters (see Table 9).
A back-of-the-envelope calculation suggests that total electricity consumption of student rooms
in the UK accounts for a very small proportion of total domestic consumption.12 On the other
hand, the share of their overall electricity consumed via the smart plug is likely to be higher than
would be the case in a regular household, so we likely capture a larger proportion of their potential
behavioral response. Most importantly, our student participants do not pay for their energy usage,
meaning we can test our incentive mechanisms in isolation (i.e. without any confounding of savings
from reduced or time-shifted energy usage). The size of the rewards in relation to total budget is
also likely to be greater for our participants than for the rest of the population.
  11 Previous  research has linked social pressure to technology adoption, adding another dimension to this channel;
we do not attempt to measure treatment or welfare effects from our channel of adoption (see Giaccherini et al.,
2019), but acknowledge that there may be an effect on adoption that, if anything, would attenuate any treatment
effects in our study.
  12 Students in Imperial residences consume on average 18kWh per week. Assuming similar consumption levels

for the approximately 650,000 student rooms in the UK, university residents’ consumption totals 11GWh per week
which, under certain assumptions, can be produced by 40 wind turbines.

                                                         5
2.2       Experimental design

To discern the likelihood of adoption based on switch-off intensity, we used a simple randomiza-
tion13 to assign prospective participants into three groups differing in the frequency and duration
of switch-off events. We implemented three cells of a 2x2 study design. The first dimension is the
number of possible switch-off events in a given 24-hour period. Some invitees were told that the
switch-off events may occur once per day, while others were told they may occur once every three
hours (i.e., up to eight times in a 24-hour period). The second dimension is the duration of the
switch-off events, which may be either 15 or 30 minutes. To bolster group-level sample size and
maximize observed switch-off events, we exclude the 15-minute switch-off once per three hours cell
of the 2x2 design, as shown in Table 1.

                                       Table 1: Treatment Group Design

                                                      3-hour intervals       24-hour intervals
                      15 minutes switch-offs                                   Treatment 1
                      30 minutes switch-offs            Treatment 3            Treatment 2
                     Notes: This table describes the three treatments with switch-off events vary-
                     ing in both duration and frequency over a 24-hour period. A 3-hour interval
                     means a switch-off event can occur once every 3 hours, i.e. up to eight times in
                     a 24-hour period.

We distributed a free smart plug to all consenting participants that we used to remotely control
electricity supply to their selected electronic device or appliance. Over the course of the study pe-
riod, we administered randomly timed switch-off events in accordance with the switch-off intensity
to which each participant had been assigned. Participants collected points based on the energy
that had been displaced through these events, as measured by the amount of energy consumed
through the plug (in Watts) immediately prior to the event multiplied by the event’s duration.14
Finally, in fortnightly intervals, we performed a lottery where participants with more points had
a higher chance of winning monetary prizes.15 Provided they accumulated at least one reward
point during the two-week period, the lottery allocated each participant with a probability to win
a voucher in proportion to their reward points.16 Table 2 shows that we distributed £550 worth
of versatile digital gift cards to 26 total winners, among which a vast majority (18) won only once,
while three participants won 4 times.

2.3       Data

Data collection began in April 2019 and lasted until mid-March 2020. In addition to recruitment
data, our dataset consists of energy consumption readings for all online POWBAL plugs at 5-
minute intervals (1,323,888 observations), the timing of the random switch-off events, and the
dates and amounts of prize money lottery winners received.

Figure 1 reports hourly figures of aggregate power consumption via the smart plugs. Out of the
231 prospective participants who signed up and received their plug, 144 went on to use it at least
once. The LSE trial started in April 2019 and ended over the summer as students successively
moved out of the halls, as reflected in the left-hand part of Figure 1. The Imperial trial initiated in
October 2019 when students moved in, and the Reading trial began in January 2020 at the start
  13 We  do not have data on characteristics of the subject pool, hence stratified randomization was not possible.
  14 To illustrate, if the power consumed by the plug prior to switch-off was 2 Watts, and the switch-off event lasted
for 30 minutes (i.e. 12 hour), the participant earned 2 Watts x 12 hour = 1 reward point. If a participant overrode
the switch-off event (i.e. by pressing the override button on the plug or via the app), we only rewarded the user for
the amount of time that the plug was switched off before the override.
  15 Our prizes came in the form of digital gift cards, redeemable in a variety of high-street and online retailers as

well as hospitality providers.
  16 The lottery worked as follows: the number of sub-lotteries for a £10 voucher was defined by the number of

participants with non-zero reward points during a given fortnight, for example 100 sub-lotteries if there had been
100 such participants. Each voucher was given away by drawing the names of participants. The probability to win
each of these draws was equal to 15% of that participant’s share of the overall number of reward points accumulated
in those two weeks.

                                                            6
Table 2: Lottery Results

                                               N     Vouchers won         Highest voucher won
                 One-time winners           18             £260                      £30
                 2-times winners            4              £120                      £50
                 3-times winners            1              £30                       £10
                 4-times winners            3              £140                      £20
                 Total                      26             £550                      £50

                 Notes: This table reports the outcomes of the fortnightly lotteries, i.e how the pay-
                 outs were distributed among lottery winners. Each line records how many participants
                 won a given number of lottery draws, the total value distributed and highest value of
                 vouchers won by x-times winners out of the entire voucher pool.

of term. The latter two trials ended on March 15, 2020 when the COVID-19 lockdown became
imminent, which led many students to move out of the halls to be with their families.17

Table 3 reports descriptive statistics on the power consumption of all connected devices (outside of
switch-off events) throughout the trials. The average power is 7.5 Watts, akin to that of a phone
charger. However, this figure hides considerable heterogeneity both between and within users.
Note that most plugs do not consume any electricity most of the time. Of the nearly 1.3 million
data measurements we observe from our active users, only 22% report non-zero consumption, while
in the remaining 78% of observations the plugs were connected to a socket without any devices
consuming energy through them. However, there are several users who at times use up to 3 kW
(about the power needed for 2 average electric space heaters).

          Table 3: Descriptive Statistics of Power Consumption via Smart Plugs (Watts)

  Statistic                            N           Mean     St. Dev.      Min      Pctl(25)      Pctl(75)      Max
  All online plugs                 1,323,888        7.53       73.10       0          0             0         3,118
  Plugs with non-zero power         288,653        34.53       153.53     3.10       6.40         25.10       3,118
  Non-zero power in week           1,235,129        8.07        75.65      0           0            0         3,118
   Notes: This table reports descriptive statistics of power consumption over the duration of the experiment. Power
  consumption is recorded in Watts. The first row reports on all plugs that are online. The second row reports on
  plugs online with non-zero consumption. The third line reports on plugs that have at least one observation with non-
  zero consumption in a given week.

Within-plug consumption changes over time as users connect different devices, turn off the outlet,
unplug the POWBAL plug, or toggle between switching a device off or on, which they can do
remotely or via a button on the plug itself. The data on power drawn when power consumption is
non-zero is reported in row 2 of Table 3, which reveals that average consumption is 34.5 Watts with
the 75th percentile at 25.1 Watts, on the order of magnitude of a small laptop or tablet charger.
This coincides with qualitative feedback from the debrief survey (see Appendix A). Figure 2 reports
the distribution of measured power for plugs with non-zero consumption.

2.4     Consumption over the course of the day

Given the importance of optimizing and integrating VPPs in increasingly smart energy systems,
demand fluctuations and VPP capacity throughout the day must be well understood. While
variation is independent of the day of the week, Figure 3 shows that the average power usage via
the plugs during different hours of the day varies substantially. During evening hours from 18:00
  17 In our pre-registry, we had indicated that trials would take place in 2018, and simultaneously. Logistical

technicalities delayed rollout at two of our sites, and the opportunity to work with the University of Reading arose
as we were preparing to implement the experiment at Imperial. Given lower than expected uptake in Imperial and
LSE, we partnered with Reading to achieve our desired sample size of 200-250, as designated in the pre-registry,
which can be found on OSF.

                                                           7
Figure 1: POWBAL Performance Over Time

                           3000           Start LSE                            Start Imperial              Start Reading     300
 Power Consumption in Wh

                                                                                                                                   Number of plugs online
                           2000                                                                                              200

                           1000                                                                                              100

                                  0                                                                                          0

                                      Apr 2019        Jul 2019             Oct 2019             Jan 2020          Apr 2020
                                                                            Time

                                                                 Energy Consumption      Plugs online

Notes: The graph reports hourly figures of aggregate power consumption in Watt hours (red curve)
and the number of plugs online (blue curve) over the duration of the experiment. Data collection
in LSE halls started in April 2019 and ended during the summer. Subsequent trials at Imperial
College London and University of Reading initiated in October 2019 and January 2020 respectively
and ended due to the emergence of COVID-19, which drew students home from March 2020.

                                                       Figure 2: Distribution of Power

                           0.08

                           0.06
 Density

                           0.04

                           0.02

                           0.00

                                      0               20                     40                    60                 80
                                                                                  Watt
Notes: The figure shows the fitted density plot of electricity consumption in Watts for non-zero
consumption over the trial period, where the cutoff at the 99th percentile is 84.3 Watts.

                                                                             8
Figure 3: Electricity Consumption Throughout the Day in Wh

          Notes: The figure depicts average electricity consumed through plugs at each hour of the day,
          reported in Watt hours. Left-hand side y-axis denotes the Wh for the sample averages, whereas
          right-hand side y-axis denotes the Wh for the UK average. The hours of the day are noted on
          the x-axis. The average hourly consumption data for the UK was adapted from Ushakova and
          Jankin Mikhaylov (2020).

(6pm) onward, consumption is more than three times higher than during the lowest consumption
period around 5:00 (5am). This pattern is comparable to within-day variations of consumption in
typical UK households (Ushakova and Jankin Mikhaylov, 2020), with the notable difference that
nighttime consumption is higher through the students’ plugs and that the morning peak is not
pronounced. Figure 3 illustrates that electricity consumption when overriding switch-off events
remains low during peak load, with the exception of 8pm. These observations signify that the
VPP’s capacity generally follows typical electricity demand patterns, so that latent and untapped
flexibility in the energy system is maximized during periods of peak load. Additionally, we see
that the load remains relatively high in late evenings, which could be important if electric vehicle
owners plug in their cars to charge at night, as incentivized in many electric vehicle energy plans.

3     Empirical Specification and Results

In this section we address our primary experimental research question of how our treatments, and
the switch-off intensity they embodied, affected the likelihood of adoption and use. To do so, we
compare adoption (a binary measure of whether participants consumed any power through the plug
at all) and use (a continuous measure of the energy consumption’s contribution toward the VPP via
the plug following adoption) across the three groups, using Treatment 2 as a reference group upon
which to compare Treatments 1 and 3. We additionally explore the impact of receiving financial
rewards on subsequent usage and rule out that participants are uniquely prosocially motivated.

3.1    Prosocial Technology Adoption: A Stylized Model

To frame our analysis, we develop a stylized model of prosocial technology adoption. POWBAL
differs from many technology adoption experiences—for instance, purchasing and using a smart
phone—in that there is no intrinsic utility for the user. The value of adoption arises exclusively
as a positive externality; e.g., in our case it facilitates increased efficiency of the electricity grid.
However, there is potentially disutility from adoption, for instance, setup costs, data sharing, and
relinquished control over select energy services in the home. As a consequence, users may need to
be compensated (Richter and Pollitt, 2018).

                                                       9
Figure 4: A Stylized Model

                                                                   Benefit from transfers (lottery)
                 Disutility/Transfer
                                                               C

                                                         Disutility from plug use

                                         A

                                B

                                                                                    Duration

               Notes: The figure models disutility from plug use (red curve) combined with benefits
               from transfers (blue curve) as a function of variations in duration of switch-off events.
               At point A, benefits form the the lottery exceed the disutility from adopting the plug.
               Conversely, at points B and C, disutility is larger than the potential benefits.

For simplicity, we assume that intensity of participation is unidimensional (e.g., frequency or
duration of a switch-off event) and that there is a fixed disutility, or ‘setup cost’, from adoption.
Moreover, with increasing switch-off intensity, the disutility of adoption increases at an accelerating
rate, as in the red line of Figure 4. We also suppose that transfer payments (e.g., lottery prizes)
increase linearly in consumption through the smart plug, as in the blue line of Figure 4. At
point B, disutility exceeds benefits and the user will not adopt. At point A, benefits are higher
than disutility so that a user will adopt. In our current setting, the user cannot choose switch-
off intensity freely; rather, intensity is exogenously allocated. Hence, depending on the shape of
the underlying functions in our experimental settings, individuals will either adopt and use the
technology (as at point A) or choose not to adopt (as at points B or C).

In the next two sections, we first analyze how the different treatments, and the disutility from
plug use that they represent, affect adoption rates. We then explore whether the rewards from the
lottery (extrinsic motivation) interact with prosocial utility (intrinsic motivation).

                                                          10
3.2     Treatment effects on adoption

3.2.1      Empirical Specification

Our treatments allow us to explore two dimensions that could have an impact on subjects’ decisions
of whether to adopt the plugs: the duration and the frequency of energy service interruption. All
else equal, we would assume that a user prefers lower frequency (i.e. a long gap between switch-off
events) and shorter duration. However, shorter and less frequent switch-off events also translate
to lower payoff likelihood, so the expected response to the different treatments is an empirical
question.

To answer this question, we run regressions of two different measures of VPP participation on the
treatments received, which vary in frequency and duration. The regressions take the following
form:
                               Yi = β0 + β1 T 1 + β2 T 3 + Hi0 γ + i                         (1)
where Yi is the measure of VPP participation. The first measure is the average electricity consumed
per week (in kWh) through individual plugs. The second measure is the percentage of weeks that
a plug is online, regardless of how much electricity is consumed through it. For example, if a user
came online for the first time in week 1, and then only every other week until the end of the trial,
the latter outcome would be 50%. T 1 and T 3 are binary variables set to 1 when a user is in a
given treatment group. The reference group is T2, where users are exogenously assigned switch-off
events of relatively long duration (30 minutes) and only once every 24 hours. Hi is a vector of
residence hall dummies.

3.2.2      The effect of varying switch-off event frequency and duration

Figure 5 graphically explores two measures of adoption—weekly energy consumption and share
of users that have their POWBAL plug online at least once during a given week—conditional on
treatment assignment. It reports the βd coefficients of the regression
                                          X
                                    Yit =      βd DU Rit + it
                                                     d

where Yit is either the weekly energy consumption (in kWh) of user i in week t (in the left panel
of the figure) or a dummy equal to 1 if a user’s plug came online in a given week (in the right
panel). DU Rd,i,t is a dummy equal to 1 if in week t user i has been participating in the trial for d
weeks. The figure reveals that energy consumption via the plug tends to be highest for Treatment
Group 2 (low frequency, long duration of switch-off event) particularly in the early weeks of the
trial. Consumption is lowest for participants assigned to the least intrusive treatment, with the
highest-intensity group in between.

In line with our discussion in section 3.1, one explanation for the low consumption in Treatment
1 may be that users do not view the disutility from switch-off events as worthwhile when prospec-
tive rewards—whether monetary or prosocial—are too low, so infrequent switch-off events lead to
user disengagement. On the other hand, frequent switch-off events inconvenience the user, so that
engagement is highest with long duration and low frequency events. Explicit regressions of con-
sumption on treatment characteristics, as per equation 1, are presented in Table 4. They support
this interpretation in that they suggest users with short duration switch-off events consume 3.3
kWh less on average over a week than the reference group, whereas high frequency reduces con-
sumption by 2.8 kWh on average, as shown in column 1. The result holds when we do not control
for the residence in which the user lives (column 2). When winsorizing the dependent variable at
1% (column 3), the magnitude and the significance of the coefficients are attenuated, but their sign
is preserved, indicating the results are not solely driven by a few outliers.18
 18 This   also holds when winsorizing at 5%. At 2% winsorization, the T1 and T3 coefficients are significant at 10%.

                                                          11
Figure 5: Treatment Effects on Plug Usage

                                         20
      Average Power consumption in kWh

                                                                                                                      80

                                                                                         Share of users online in %
                                         15

                                         10

                                                                                                                      40

                                          5

                                          0
                                                                                                                      0

                                         −5
                                              0     10            20             30                                        0          10           20        30
                                                  Trial Duration in weeks                                                          Trial Duration in weeks

                                                                   Treatment 1        Treatment 2                              Treatment 3

     Notes: The left-hand side graph reports the average energy consumption in kWh consumed through the plugs
     for each treatment group across all sites over the trial duration measured in weeks where time 0 is when the
     user first uses the plug. The right-hand side graph reports the share in percent of users that come online at
     least once, even briefly, during a given week for each treatment group across all sites over the trial duration in
     weeks.

Additional results support this interpretation. The gap between Treatment Group 2 and the other
groups becomes initially bigger as the trial proceeds. This is consistent with participants learning
about their rewards: users in group 2 respond to higher rewards whereas users in group 1 lose
interest because of their lower chance of receiving rewards. We also see that users in group 2 are,
toward the end of the trial period, the most likely to remain engaged, leaving their plugs online
for more weeks, as shown in the right panel of Figure 5. This effect is less pronounced than the
consumption effect: in the last column of Table 4, we find that users with a shorter switch-off
duration treatment are 7 percentage points less likely to continue participating compared to the
reference group, though the effect is not statistically significant.

There are a number of caveats to this interpretation. First, Figure 5 illustrates powerfully that
there is non-trivial attrition. By week 15 of the trial most users have abandoned adoption. It
is however noteworthy that the observed drop is conflated by the end of term (in the LSE case)
and the onset of the COVID-19 lockdown (in the Reading case), as illustrated by Figure 6, which
repeats Figure 5 for our three trial sites separately. Likewise, 60% of debrief survey respondents
expressed an interest in remaining involved in POWBAL initiatives in the future. Note that for
the Imperial site where we had about 27 weeks of data before the lockdown began, attrition is
much lower.

Interestingly, the relative performance of the treatment groups is different for the Imperial site
compared to the others. Hence, we have to be cautious in drawing overly strong conclusions from
the aggregated results. However, consistent with earlier results, group 1 comes lower in terms of
participation likelihood toward the end of the trial. Hence, we can conclude with some confidence
that the need to provide sufficient incentive to participate will dominate, within reason, factors
such as the inconvenience of having plugs switch off too often or for too long.

                                                                                      12
Figure 6: Treatment Effects on Plug Usage by Site

                                                                               LSE
                                                                                                                   120
 Average Power consumption in kWh

                                    15

                                                                                      Share of users online in %
                                    10                                                                              80

                                     5                                                                              40

                                     0
                                                                                                                        0

                                         0   10                  20           30                                            0          10                  20    30
                                             Trial Duration in weeks                                                                   Trial Duration in weeks

                                                                              Imperial
 Average Power consumption in kWh

                                                                                      Share of users online in %

                                    10                                                                             100

                                     5                                                                              50

                                     0
                                                                                                                        0

                                    −5
                                         0   10                  20           30                                            0          10                  20    30
                                             Trial Duration in weeks                                                                   Trial Duration in weeks

                                                                              Reading
                                    40
 Average Power consumption in kWh

                                                                                      Share of users online in %

                                    30                                                                             90

                                    20                                                                             60

                                    10
                                                                                                                   30

                                     0
                                                                                                                    0

                                         0   10                  20           30                                            0         10                  20     30
                                             Trial Duration in weeks                                                                  Trial Duration in weeks

                                                                Treatment 1        Treatment 2                                  Treatment 3

Notes: The left-hand side graph reports the average energy consumption in kWh consumed through the plugs
for each treatment group across the various sites over the trial duration measured in weeks. The right-hand
side graph reports the share in % of users that come online at least once during a given week for each treatment
group across LSE sites over the trial duration in weeks.

                                                                                   13
Table 4: Regressions on Treatment Characteristics

                                  Average Weekly Consumption (kWh)
                                     (1)             (2)                (3)            Users Online
                                           ∗               ∗∗                  ∗
Short duration (T1)               −3.272         −3.289              −1.659                −7.036
                                  (1.671)         (1.628)            (0.944)               (4.942)

High Frequency (T3)               −2.768∗         −2.703∗            −1.327                −4.432
                                  (1.633)         (1.598)            (0.927)               (4.850)

Reference Group (T2)              4.809∗∗         4.874∗∗∗           3.124∗∗∗            40.682∗∗∗
                                  (1.848)          (1.176)            (0.682)             (3.570)

Observations                        144              144                144                  144
Site Controls                       Yes              No                 No                   No
Winsorized                          No               No                 Yes                  No
R2                                 0.033            0.032              0.024                0.014
Adjusted R2                        0.005            0.018              0.010               0.0004
Residual Std. Error                7.853            7.802              4.524               23.679
F Statistic                        1.182            2.305              1.716               1.031
 Notes: These estimates are the result of OLS regressions. The table reports the effect of varying the
duration and frequency of switch-off events on the weekly average electricity consumption in kWh, as
well as the number of users whose plugs are online. Each coefficient represents the difference relative to
the reference group so as to estimate the treatment effects of long duration and high frequency switch-
off events. The dependent variable is average weekly electricity consumption in columns 1 through 3.
In column 1, a set of dummies controls for the student residence in which the participant resides. In
column 3, the average weekly electricity consumption is winsorized at 1%. In column 4, the dependent
variable is the percentage of weeks that a given user was online at least once after their first connec-
tion. Robust standard errors are reported in parentheses. Significance levels are indicated as ∗ p
3.3     Participation motives

Despite the disutility from interruption of energy services, at least two advantages may drive users
to adopt. A purely prosocial motivation may arise since POWBAL technology promises to enable
clean energy generation, and users may desire to win lottery prizes.19 If users are exclusively
driven by prosocial motives, they should not respond to lottery wins by increasing their energy
consumption through the plug to increase their chances of future wins. However, if they are driven
solely by private rewards and if they update their expectations about those rewards according
to observed lottery wins, we would expect to find a relationship between receiving rewards and
participation intensity. In other words, a boost in VPP contributions in response to rewards being
paid out is a sufficient condition for the presence of reward effects (although not a necessary one).

We develop a simple model to illustrate this phenomenon. Suppose users’ utility is defined as

                                          U = (κri + si ) ei − eηi − γ

where ei is participation intensity, ri represents the financial reward, κ is the importance of rewards
to users, si is the prosocial benefit, and γ is the fixed setup cost. We assume that η > 1, which
implies increasing marginal disutility from higher participation intensity.

Hence optimal participation is defined by the first-order condition

                                          U 0 = κri + si − ηeη−1
                                                             i   =0

so that optimal participation intensity becomes

                                                                            1
                                                                           η−1
                                                               κri + si
                                                 e∗i   =
                                                                  η

If there is uncertainty over rewards, consumers will maximize expected utility E{U } which implies
that optimal adoption becomes

                                                                               1
                                                                              η−1
                                                           κEit {ri } + si
                                          e∗it   =
                                                                 η

with Eit {ri } being users’ expectation about rewards at time t. If we find that actual reward
payouts in the past (e.g., rit−1 ) drives current participation eit , we can infer that users form their
expectations about payouts on the basis of what they have actually received in the past:

                                                  Eit {ri } = f (rit−1 )

Hence, past payouts become an instrument for the effect of rewards on participation, though we
do not directly observe Eit {ri }. We can then write the reduced form effect in this setting as

                                                                          2−η
                            ∂eit     1                 κEit {ri } + si      η−1
                                                                                  κ ∂f (rit−1 )
                                  =                                                 ×
                           ∂rit−1   η−1                      η                    η   ∂rit−1

Consequently, a regression of eit on rit−1 will provide the response of expectations to rewards
times the marginal response to rewards. For an entirely prosocially driven user (with κ = 0) this
  19 Of 52 debrief survey responses, 28 (54%) claimed environmental benefits as their primary motive for participa-

tion, whereas 22 (42%) and 2 (4%) claimed that money and free technology were their primary motives, respectively.

                                                                 15
marginal response would be zero. For users that are not exclusively prosocially driven, it seems
plausible to assume that

                                                                                               ∂f (rit−1 )
                                                                                       0<                  β
                                                                                 η−1             η                   η

Figure 7 plots the change in weekly energy consumption against the change in reward payments
received. We find large variation in both variables. The figure also reveals a positive relationship
between energy consumption at t and rewards at t − 1, which suggests that more reward payments
will encourage participants to increase the amount of electricity they consume through the plugs.

                                                                        Figure 7: Effect of Change in Rewards on Usage
           Change in weekly energy consumption (Arc Growth Rate)

                                                                    2

                                                                    1

                                                                    0

                                                                   −1

                                                                   −2

                                                                           −20                          0                   20
                                                                                       Change in reward t−1 (in £)
          Notes: The figure shows a scatter plot between percentage changes in weekly energy consumption
          and lagged changes in reward payouts (t-1). We see that a reward payout (in £) in the previous
          week (t-1) is associated with an increase in energy consumption via the plug.

We examine this response in greater detail by running regressions of the form 3.3 where we also
assume that there is a user-specific fixed effect:

                                                                                             it = αi + νit

To deal with unobservable user-specific variation we run the regression in first differences. Table
5 presents the results. Column 1 suggests that a reward payout in the previous week will increase

                                                                                                   16
consumption in the current week by 51Wh per £1 paid out. Column 2 includes both the rewards
lagged at t − 1 and t − 2, which suggests that the consumption increase remains positive and
significant two weeks after a payout. However, we also see that this additional boost disappears
by the third week after receiving payout. Hence, our first column estimate seems to be a good
representation of the net effect.

                Table 5: Regressions of Energy Consumption on Reward Payouts

                                                           Dependent variable:
                                             Change in weekly energy consumption
                                                (1)               (2)               (3)
                                                      ∗∗                ∗∗
                   Reward Week -1            52.404             64.020          86.218∗
                                             (20.818)           (25.129)        (50.468)

                   Reward Week -2                               48.868∗∗        87.548∗∗
                                                                (23.507)        (39.620)

                   Reward Week -3                                −7.469          −7.492
                                                                (16.817)        (44.639)

                   Wh Week -1                                                    −0.131
                                                                                 (0.347)

                   Wh Week -2                                                     0.095
                                                                                 (0.230)

                   Wh Week -3                                                     0.100
                                                                                 (0.283)

                   Week controls               Yes               Yes               Yes
                   Observations                1458              1458              1458
                   Users                        144               144               144
                    Notes: These estimates are the result of OLS regressions. The dependent
                   variable is the change in weekly electricity consumption in Wh from week t-
                   1 to week t. The explanatory variables take the value 1 if the user received a
                   reward in week t-1, t-2 or t-3, and zero otherwise. The table reports the ef-
                   fect of receiving a reward in previous weeks on the change in electricity con-
                   sumption. Column 3 also includes as controls weekly electricity consump-
                   tion in weeks t-1, t-2 and t-3. Robust standard errors are reported in paren-
                   theses. Significance levels are indicated as ∗ p
average consumption. Of course, rewards need to be determined in relation to marginal increases
in either revenue or carbon reductions, which will be explored below.

3.4    Overriding

Our discussion thus far has focused on VPP participation and participation intensity, which are the
primary outcomes of interest in our study. We explore a further dimension of participant response
to the exogenous switch-off schedule: user overriding behavior (i.e. canceling the switch-off event
after it has begun). Table 6 reports regressions where the dependent variable is a binary variable
equal to 1 if a switch-off event is overridden by the user, and the independent variable of interest
is treatment assignment. In total we observe 6,725 switch-off events.

At first glance, treatment assignment (i.e. switch-off intensity) appears to have little impact on
overriding behavior—with overrides occurring during 5.5% of switch-off events in Treatment 2
(Reference Group) with a statistically insignificant effect of changes to variation and frequency.
This low figure is due in part to many of the switch-off events occurring during periods where users
are not consuming power via the plugs. Hence, in column 2, we only consider overrides in cases
where a switch-off event occurs after at least five-minutes of non-zero plug load (in Watts). In this
sample, overriding occurs in 7% to 17% of cases. Comparing the variation in overriding behavior
between treatment groups leads to clear results: overriding is most prevalent in Treatment 2 with
relatively long but infrequent switch-off events. On the other hand, the amount of overriding
in Treatments 1 and 3 are very similar. One might have expected most overriding to occur in
Treatment 3, wherein participants experience both long and frequent switch-off events. This result
could be driven by users connecting different types of devices under the various treatments. Longer
treatments may mean users ensure ex ante that any energy services from connected devices are
time flexible and therefore better suited for withstanding switch-off events.

Columns 3 and 4 explore whether there is a relationship between pre-switch-off energy consumption
and overriding. A positive relationship would be of concern as this would imply that the overriding
behavior might substantially reduce the capacity of the virtual power plant. Column 3 includes the
whole sample whereas in column 4 we only include the switch-off events with positive pre-switch-
off power usage (as in column 2). The strong association between power usage and overriding
in column 3 is not present in column 4. Hence, the effect in column 3 is entirely driven by the
extensive margin of power users with positive usage.

3.5    Optimal operation and value of load balancing

The primary benefit of this load balancing technology adopted at scale would be system-wide cost
reductions, which will increase substantially as larger shares of electricity derive from intermittent
renewable generation (see Strbac, 2008). Nevertheless it is insightful to benchmark the performance
of our emerging virtual power plant against the current performance of the UK electricity market.
To do so, we need to determine the most efficient way to operate the plant; that is, we need to
optimize the demand response automation by taking into account individuals’ usage patterns as
well as fluctuations in cost and/or carbon content of electricity generation over time.

Such optimization is not trivial as constraints on plug usage give rise to a real options problem:
switching off a plug in period t will remove the option to switch off the plug for a number of
subsequent periods. We develop an algorithm that accounts for this issue as well as for market data
and user-specific behavioral data from our field experiments. We describe the highly personalized
algorithm for plug control that results from it as “power doctoring”. To derive the algorithm and
simplify its exposition, we assume in what follows that an individual plug is under treatment 3,
i.e. switch-off events last 30 minutes within 3 hourly (or 6 half-hourly) intervals, and that we are
solving it for a week. The algorithm defines at any time t whether the plug should be switched
off (sot = 1) or not (sot = 0). θ(t) characterizes which half-hour of the week t is at the start of,
and therefore takes in this case values 1 to 336. We define the number of intervals (here, 336)

                                                 18
You can also read